Create the Baseline Data Series

The trading strategy will be compared to the S&P 500, shown below. We use this to create an S&P 500 rate of change series. The trading rule will move in and out of this series.

Perform the Backtesting and Plot the Results

In this final analysis step the trading rule is plotted along with the indicator in the top pane. The middle pane shows how the trading rule modified the rate of change series. The bottom plots shows how the investment performed, compared to the S&P 500.

df.data$retRec <- df.data$retBase * df.data$RecInitTrade
df.symbols <-
  rbind(
    df.symbols,
    data.frame(
      string.symbol = "retRec",
      string.source = "Calc",
      string.description =  "Rate of Change, Recession Initiation Rule",
      string.label.y = "Percent",
      float.expense.ratio = -1.00,
      Max030 = FALSE,
      Max180 = FALSE,
      date.series.start = dt.start.prediction,
      date.series.end = as.Date(Sys.Date())
    )
  )

df.data$eqRec <- exp(cumsum(df.data$retRec))
df.data$eqRec <- df.data$eqRec/df.data[min(which(df.data$date>dtStartBackTest)),"eqRec"]
df.symbols <-
  rbind(
    df.symbols,
    data.frame(
      string.symbol = "eqRec",
      string.source = "Calc",
      string.description =  "Equity Return, Recession Initiation Rule",
      string.label.y = "$1 Invested",
      float.expense.ratio = -1.00,
      Max030 = FALSE,
      Max180 = FALSE,
      date.series.start = dt.start.prediction,
      date.series.end = as.Date(Sys.Date())
    )
  )

dataTrade <- "RecInitTrade"
dataRet <- "retRec"
dataEq <- "eqRec"

ylimBackTest = c(0, 250)
p1 <-
  plotBack(
    dfRecession,
    df.data,
    dataTrade,
    dataRet,
    dataEq,
    dfPred,
    bOverlay = TRUE,
    dtStartBackTest,
    ylimBackTest
  )

The trading strategy resulted in an improved return, although most of this comes after the 2007-2008 recession. In that recession the trading rules return to the long position at the exact market bottom. This is in contrast to the 1999-2000 recession where the trading rule returns to long before the market bottoms.

S&P 500 SMA Trends

Take a look at recent activity in the equities market. This section looks at the S&P 500 close value to the 50 day and 200 day simple moving average (SMA).

This is a longer view for the S&P 500 trend.

S&P 500 Normalized by GDP

The last two years compare favorably with the period around the late 1950’s. Need to dig into this one.

datay <- "GDPSP500"
ylim <- c(0.10, 0.16)
my.data <- plotSimilarPeriods(df.data, dfRecession, df.symbols, datay, ylim, i.window = 730)
my.data[[1]]

Look at how the different segments of the market move

datay <- "GSPC.CloseBYMDY.Close"
ylim <- c(0, 20)
dtStart = as.Date('1980-01-01')
plotSingle(dfRecession, df.data, "date", datay, getPlotTitle(df.symbols, datay), "Date", 
            getPlotYLabel(df.symbols, datay), c(dtStart, Sys.Date()), ylim, TRUE)

S&P 500 Normalized moving average

Look at moving average relationship by dividing the S&P 500 open price by the 200 day SMA.

datay <- "GSPC.Open_mva200_Norm"
ylim <- c(50, 125)
dt.start = as.Date('2008-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dt.start)

Crossovers

Look at the 50 DMA versus 200 DMA, often used as a technical indicator of market direction.

datay <- "GSPC.Open_mva050_mva200"
ylim <- c(-200, 200)
plotSingleQuick(dfRecession, df.data, datay, ylim, dtStartBackTest)

datay <- "GSPC.Open_mva050_mva200_sig "
ylim <- c(0.0, 1.0)
plotSingleQuick(dfRecession, df.data, datay, ylim, dtStartBackTest)

df.data$ret050MAMinus200MA <- df.data$retBase * df.data$GSPC.Open_mva050_mva200_sig
df.data$ret050MAMinus200MAShort <- df.data$retBase * abs(df.data$GSPC.Open_mva050_mva200_sig-1.0)

df.symbols <-
  rbind(
    df.symbols,
    data.frame(
      string.symbol = "ret050MAMinus200MA",
      string.source = "Calc",
      string.description =  "Rate of Change, 50 DMA - 200 DMA Rule",
      string.label.y = "Percent",
      float.expense.ratio = -1.00,
      Max030 = FALSE,
      Max180 = FALSE,
      date.series.start = dt.start.prediction ,
      date.series.end = as.Date(Sys.Date())
    )
  )

df.data$ret050MAMinus200MARet <- exp(cumsum(df.data$ret050MAMinus200MA))
df.data$ret050MAMinus200MARet <- df.data$ret050MAMinus200MARet/df.data[min(which(df.data$date>dtStartBackTest)),"ret050MAMinus200MARet"]
df.symbols <-
  rbind(
    df.symbols,
    data.frame(
      string.symbol = "ret050MAMinus200MARet",
      string.source = "Calc",
      string.description =  "Equity Return, 50 DMA - 200 DMA Rule",
      string.label.y = "$1 Invested",
      float.expense.ratio = -1.00,
      Max030 = FALSE,
      Max180 = FALSE,
      date.series.start = dt.start.prediction ,
      date.series.end = as.Date(Sys.Date())
    )
  )


dataTrade <- "GSPC.Open_mva050_mva200_sig"
dataRet <- "ret050MAMinus200MA"
dataEq <- "ret050MAMinus200MARet"
p1 <-
  plotBack(
    dfRecession,
    df.data,
    dataTrade,
    dataRet,
    dataEq,
    dfPred,
    bOverlay = FALSE,
    dtStartBackTest,
    ylimBackTest
  )

S&P 500 TTM P/E

Market prices can out-run earnings so take a look at price to earnings.

Focus on some of the more recent activity

S&P 500 Sales

datay <- "MULTPLSP500SALESQUARTER"
ylim <- c(500, 1500)
dt.start <- as.Date('1999-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dt.start)

datay <- "MULTPLSP500SALESQUARTER"
ylim <- c(500, 1500)
dt.start = as.Date('2001-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dt.start)

Unit Profits

The series peaks in the middle of a bull market.

S&P 500 dividends

12-month real dividend per share inflation adjusted November, 2018 dollars. Data courtesy Standard & Poor’s and Robert Shiller.

https://www.quandl.com/data/MULTPL/SP500_DIV_MONTH-S-P-500-Dividend-by-Month

datay <- "MULTPLSP500DIVMONTH_YoY"
ylim <- c(-50, 50)
dtStart = as.Date('1910-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dtStart, b.percentile = FALSE)

datay <- "MULTPLSP500DIVMONTH"
ylim <- c(20, 60)
dtStart = as.Date('2001-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dtStart, b.percentile = FALSE)

datay <- "MULTPLSP500DIVMONTH_YoY"
ylim <- c(-40, 20)
dtStart = as.Date('2001-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dtStart, b.percentile = FALSE)

S&P 500 dividend yield (12 month dividend per share)/price. Yields following September 2018 (including the current yield) are estimated based on 12 month dividends through September 2018, as reported by S&P. Sources: Standard & Poor’s for current S&P 500 Dividend Yield. Robert Shiller and his book Irrational Exuberance for historic S&P 500 Dividend Yields.

https://www.quandl.com/data/MULTPL/SP500_DIV_YIELD_MONTH-S-P-500-Dividend-Yield-by-Month

datay <- "MULTPLSP500DIVYIELDMONTH"
ylim <- c(0, 12)
dtStart = as.Date('1910-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dtStart, b.percentile = FALSE)

datay <- "MULTPLSP500DIVYIELDMONTH"
ylim <- c(1, 4)
dtStart = as.Date('2001-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dtStart, b.percentile = FALSE)

S&P 500 Volume

The log of the S&P volume has some interesting patterns, but nothing that seems to help with a recession indicator.

That is one spiky data series. Not sure there is a lot to help us here.

Russell 2000

Take a look at recent activity in the small cap market.

datay <- "RLG.Open"
datay.aux <- "RLG.Open_mva200"
datay.title <- getPlotTitle(df.symbols, datay.aux, str.sep = "\n")
datay.aux.1 <- "RLG.Open_mva050"
datay.title.1 <- getPlotTitle(df.symbols, datay.aux.1, str.sep = "\n")
ylim <- c(1000, 1750)
myPlot <-
  plotSingle(
    dfRecession,
    df.data,
    "date",
    datay,
    getPlotTitle(df.symbols, datay),
    "Date",
    getPlotYLabel(df.symbols, datay),
    c(dt.recent, Sys.Date()),
    ylim,
    b.legend = TRUE,
    b.percentile = FALSE,
    b.long.legend = TRUE
  )
myPlot <-
  myPlot + geom_line(
    data = df.data,
    aes_string(
      x = "date",
      y = datay.aux,
      colour = shQuote(datay.title)
    ),
    na.rm = TRUE
  )
myPlot + geom_line(
  data = df.data,
  aes_string(
    x = "date",
    y = datay.aux.1,
    colour = shQuote(datay.title.1)
  ),
  na.rm = TRUE
)

S&P 500 to Rusell 2000 Correlation

datay1 <- "RLG.Open"
ylim1 <- c(0, 2000)

datay2 <- "GSPC.Open"
ylim2 <- c(0, 3000)

dtStart <- as.Date("1jan2003","%d%b%Y")

w <- 30
corrName <- calcRollingCorr(dfRecession, df.data, df.symbols, datay1, ylim1, datay2, ylim2, w, dtStart)

## Warning in max.default(structure(numeric(0), class = "Date"),
## structure(numeric(0), class = "Date"), : no non-missing arguments to max;
## returning -Inf

## Warning in min.default(structure(numeric(0), class = "Date"),
## structure(numeric(0), class = "Date"), : no non-missing arguments to min;
## returning Inf

S&P 500 to MDY (Mid-cap) 2000 Correlation

datay1 <- "RLG.Open"
ylim1 <- c(0, 2000)

datay2 <- "MDY.Open"
ylim2 <- c(0, 500)

dtStart <- as.Date("1jan2003","%d%b%Y")

w <- 30
corrName <-
  calcRollingCorr(dfRecession,
                  df.data,
                  df.symbols,
                  datay1,
                  ylim1,
                  datay2,
                  ylim2,
                  w,
                  dtStart)

## Warning in max.default(structure(numeric(0), class = "Date"),
## structure(numeric(0), class = "Date"), : no non-missing arguments to max;
## returning -Inf

## Warning in min.default(structure(numeric(0), class = "Date"),
## structure(numeric(0), class = "Date"), : no non-missing arguments to min;
## returning Inf

Dividend Stocks

This is an interesting series, they should perform better through the recessions. Unfortunately they are short lived so there is not much data so this is more of a place holder for now.

datay <- "NOBL.Open"
ylim <- c(40, 75)
dt.start <- as.Date('2014-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dt.start)

NYSE Margin Debt

Taking a look at margin debt. NYXDATA stopped providing NYSE margin debt data on Dec 2017. Data is available from FINRA, but it includes more accounts than the data did for NYXdata. I stitched togeter the data sets: data after Jan 2010 include NYSE+Others, data prior is just NYSE account data scaled up to match the FINRA data.

It tends to creep up when there is a frenzy in the stock market.

datay <- "FINRAMarginDebt_Log"
ylim <- c(5, 15)
plotSingleQuick(dfRecession, df.data, datay, ylim)

Take a close look at recent activity

datay <- "FINRAMarginDebt"
ylim <- c(100000, 800000)
dt.start <- as.Date('2000-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dt.start)

Sometimes it is more helpful to view year over year growth.

datay <- "FINRAMarginDebt_YoY"
ylim <- c(-100, 50)
dt.start <- as.Date('1960-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dt.start, b.percentile = TRUE)

More near-term trend.

datay <- "FINRAMarginDebt_YoY"
ylim <- c(-100, 75)
dt.start <- as.Date('2000-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dt.start)

OCC Options Volumes

See what is happening with the options volumes for equities. (From: https://www.theocc.com/webapps/historical-volume-query)

Looks like options on non-equity co-occurs with peaks/troughs?.

CBOE VIX

As markets become complacent (low VIX) and high values, peaks often occur.

Compare the VIX to some of the ETF’s out there.

There

Not much predictive in VIX, take a quick look at the smoothed derivative.

S&P Daily Swings

Daily changes in the S&P should correlate well with the VIX.

More of a correlating series than a predictor.

Unemployment rates

Unemployment rates will probably be useful, let’s take a look at the U-3. The data is a little noisy so there is also a smoothed version plotted. There seems to be a relationship between the unemployment rate and the recessions, but it could be a lagging indicator. This will be explored a little bit more later.

Looking at the unemployment rate, the eye is drawn to the rise and fall of the data, this suggests that the derivative might be helpful as well. The figure below shows the results, using a Savitzky-Golay FIR filter. It looks like the unemployment rate peaks in the middel of the recession. That peak might be a good buy signal.

Unemployment rates, year-over-year

Both the headline unemployment and U-6 number changes are similar. During the upswing on the cycle it does look like the headline number falls faster than U-6

The second derivative of the unemployment rate does have zero crossings near the middle point of a recession. This would make it a helpful buy signal for the trading strategy.

Unemployment rates, similar periods

Historically the last two years of record low unemployment appear most similar to the 1971-1973 time frame. Just before inflation took off.

datay <- "UNRATE"
ylim <- c(3.3, 4.5)
i.window = 730
my.data <- plotSimilarPeriods(df.data, dfRecession, df.symbols, datay, ylim, i.window)
my.data[[1]]

Unemployment rates, U-6 and headline number.

Let’s also take a look at the total unemployed, U-6. It continues to fall as the headline number stabilizes as people return to the work force. An indicator the cycle is beginning to top out.

Difference between U6 and U3 to see how close the economy is getting to full employment.

Initial jobless claims

We will also take a look at initial jobless claims, this should start to rise just before the unemployment rate.

datay <- "ICSA"
ylim <- c(100000, 700000)
dt.start <- as.Date('1968-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dt.start)

It looks like the jobless claim tend to peak more towards the end of the recession. It does not seem to be as strong of a sell indicator as the U-3 rate.

datay <- "ICSA_SmoothDer"
ylim <- c(-1500, 1500)
dt.start <- as.Date('1968-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dt.start, b.percentile = TRUE)

Jobless claims have a seasonal component to them. One way to reduce this effect is to calculate year over year growth. That helps some, the peaks seem to be more closely aligned with the middle to end of recessions.

datay <- "ICSA_YoY"
ylim <- c(-75, 75)
dt.start <- as.Date('1968-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dt.start, b.percentile = TRUE)

Take a closer look at recent data

datay <- "ICSA_YoY"
ylim <- c(-50, 50)
dt.start <- as.Date('2000-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dt.start, b.percentile = TRUE)

Take a look at the percentage of the population looking for work

datay <- "LNU03000000BYPOPTHM"
ylim <- c(0, 6)
datay_aux <- "UNEMPLOYBYPOPTHM"
dt.start <- as.Date('1968-01-01')
my.plot <- plotSingle(dfRecession, df.data, "date", datay, getPlotTitle(df.symbols, datay), "Date", 
            getPlotYLabel(df.symbols, datay), c(dt.start, Sys.Date()), ylim, TRUE)
my.plot + geom_line(data=df.data, aes_string(x="date", y=datay_aux, colour=shQuote(datay_aux)), na.rm = TRUE)

A bit more recent trend

datay <- "LNU03000000BYPOPTHM"
ylim <- c(0, 6)
datay_aux <- "UNEMPLOYBYPOPTHM"
my.plot <- plotSingle(dfRecession, df.data, "date", datay, getPlotTitle(df.symbols, datay), "Date", 
            getPlotYLabel(df.symbols, datay), c(dt.recent, Sys.Date()), ylim, TRUE)
my.plot + geom_line(data=df.data, aes_string(x="date", y=datay_aux, colour=shQuote(datay_aux)), na.rm = TRUE)

Unemployment Level

datay <- "UNEMPLOY"
ylim <- c(-75, 20000)
dt.start <- as.Date('1968-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dt.start)

datay <- "UNEMPLOY_YoY"
ylim <- c(-25, 50)
dt.start <- as.Date('1968-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dt.start)

Payrolls

Look at the BLS data on payrolls. Check the NSA series, then we will look at YoY data.

datay <- "PAYNSA"
datay_aux <- "PAYNSA_Smooth"
ylim <- c(20000, 160000)
b.legend <- TRUE
b.percentile <- FALSE
myPlot <- plotSingle(dfRecession, df.data, "date", datay, getPlotTitle(df.symbols, datay), "Date", 
            getPlotYLabel(df.symbols, datay), c(as.Date("1jan1950","%d%b%Y"), Sys.Date()), ylim, b.legend, b.percentile)
myPlot + geom_line(data=df.data, aes_string(x="date", y=datay_aux, colour=shQuote(datay_aux)), na.rm = TRUE)

datay <- "PAYNSA_YoY"
ylim <- c(-7.5, 7.5)
b.legend <- TRUE
b.percentile <- TRUE
plotSingle(dfRecession, df.data, "date", datay, getPlotTitle(df.symbols, datay), "Date", 
            getPlotYLabel(df.symbols, datay), c(as.Date("1jan1950","%d%b%Y"), Sys.Date()), ylim, b.legend, b.percentile)

datay <- "PAYNSA_YoY"
ylim <- c(-7.5, 7.5)
b.legend <- TRUE
b.percentile <- FALSE
dt.start <- as.Date('2000-01-01')
plotSingle(dfRecession, df.data, "date", datay, getPlotTitle(df.symbols, datay), "Date", 
            getPlotYLabel(df.symbols, datay), c(dt.start, Sys.Date()), ylim, b.legend, b.percentile)

Hours worked

Sparked by an article at Mises (https://mises.org/wire/how-alexandria-ocasio-cortez-misunderstands-american-poverty), take a look at average weekly hours

datay <- "CEU0600000007"
ylim <- c(36, 43)
plotSingle(dfRecession, df.data, "date", datay, getPlotTitle(df.symbols, datay), "Date", 
            getPlotYLabel(df.symbols, datay), c(as.Date("1jan1945","%d%b%Y"), Sys.Date()), ylim, TRUE)

The time series is pretty lumpy, plot the YoY change

datay <- "CEU0600000007_YoY"
ylim <- c(-7.5, 7.5)
b.percentile <- TRUE
dt.start <- as.Date('1968-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dt.start, b.percentile)

A more recent look

datay <- "CEU0600000007_YoY"
ylim <- c(-5, 5)
b.percentile <- FALSE
dt.start <- as.Date('2010-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dt.start, b.percentile)

Retail sales, aggregate

Retail sales also change during recession. As the plot below shows, it seems to follow the trend of industrial production. It might be too strongly correlated to add much to the model. The will be examined in the correlation section.

The derivative of retail sales is a little more erratic than is was the industrial products. Looks like it might be helpful to include in the model as well.

Retail sales, aggregate year-over-year

Take a look at year-over-year changes

Retail sales and unemployment correlations

Let’s see how that looks on year over year basis. Interesting to compare to unemployment rates there appears to a correlation over the long term.

## Scale for 'y' is already present. Adding another scale for 'y', which
## will replace the existing scale.

There is some similarity. The rolling correlation shows the inverse relationship prior to a recession.

datay1 <- "RSALESAGG_YoY"
ylim1 <- c(-12.5, 7.5)

datay2 <- "UNEMPLOY_YoY"
ylim2 <- c(-30, 50)

dtStart <- as.Date("1jan1970","%d%b%Y")

w <- 180
corrName <- calcRollingCorr(dfRecession,df.data, df.symbols, datay1, ylim1, datay2, ylim2, w, dtStart)

Retail sales correlation and industrial production

Industrial production and retail sales look very similar so the plot below shows the 360 correlation. The corerlation does tend to fall around a recession, although 2008 was so bad that they both fell together. Not sure if it is that useful.

datay1 <- "INDPRO"
ylim1 <- c(40, 125)

datay2 <- "RSALESAGG"
ylim2 <- c(100000, 200000)

dtStart <- as.Date("1jan1981","%d%b%Y")

w <- 30
corrName <- calcRollingCorr(dfRecession, df.data, df.symbols, datay1, ylim1, datay2, ylim2, w, dtStart)

It is interesting to see the strong correlation; however, I suspect this is due to more to the shape of the trends. How do the YoY correlations look? They are a little less correlated, probably better to use in the machine learning later.

datay1 <- "INDPRO_YoY"
ylim1 <- c(-20, 20)

datay2 <- "RSALESAGG_YoY"
ylim2 <- c(-20, 20)

dtStart <- as.Date("1jan1981","%d%b%Y")

w <- 30
corrName <- calcRollingCorr(dfRecession, df.data, df.symbols, datay1, ylim1, datay2, ylim2, w, dtStart)

Advance Retail Sales

This is an advanced estimate of the retail sales value.

Also take a look at year over year

Retail sales and the labor market

Real Personal Income

Real Personal Income (Excluding Transfer, Annual)

During a recession real personal income falls. In the plot the peaks can be seen prior to each recession.

datay <- "W875RX1"
ylim <- c(0, 15000)
plotSingleQuickModern(datay, ylim)

The features we are interested in are the peaks and valleys so we’ll use the derivative to get to those. Interesting, there is usually a first zero crossing before a recession and a second during or just after the recession.

Real personal income might have some seasonal variance, but it seems the year over year change tells the same story.

Consumer price index

What does CPI look like?

datay <- "CPIAUCSL"
ylim <- c(0, 300)
plotSingleQuickModern(datay, ylim)

Check out the YoY growth

datay <- "CPIAUCSL_YoY"
ylim <- c(-2, 13)
plotSingleQuickModern(datay, ylim)

CPI to PPI

Suggested by Charlie, it can be helpful to look at the relationship between producer prices and consumer prices.

## Scale for 'y' is already present. Adding another scale for 'y', which
## will replace the existing scale.

Producer Price Index (Commodities)

Crude oil

Look at a trend of West Texas Intermediate (WTI)

Take a look at both WTI and Brent crude.

Gold

As risks increase investors often flock to safe haven assets like gold. An up-tick in prices can indicate investor uncertainty. This can be seen in the nominal price plot around 1980 and again in 2007.

This plots out the real price of gold by two different deflators. PPI corrected price is a little higher, to be expected since CPI also includes the effects of sales tax and imports. The spike in 1980 is especially pronounced in this series.

See how nominal and real prices look year over year. From the long-term view seems like there is little difference in the three series. Although not shown, even over the near-term there is little difference in the series.

See how gold correlates with the VIX. Both gold and VIX should respond to investor axiety, but it doesn’t look like it correlates very well.

## Warning: Removed 126 rows containing non-finite values (stat_smooth).

Copper

Dr. Copper has a reputation as an indicator of economic malaise, but it does not seem to have much of a correlation with the recessions. The series below is from CME via Quandl. It has a lot of data so I am also looking at the smoothed version.

Copper is one of the commodities in the PPI so it is a bit of a proxy for how copper is doing relative to the basket of commodities.

The change in prices, year over year, do generally peak prior to a recession. The time and shape of this peak varies, but it still might be helpful. A couple of the large troughs do seem to correlate with the end of the recession. Likely this is because industrial production has also fallen.

There is some correlation between copper and the smooth recession initiator, especially at the end of the recession.

datay1 <- "INDPRO_YoY"
ylim1 <- c(-25, 25)

datay2 <- "CHRISCMEHG1_YoY"
ylim2 <- c(-160, 60)

dtStart = as.Date('1960-01-01')

w <- 360
corrName <- calcRollingCorr(dfRecession, df.data, df.symbols, datay1, ylim1, datay2, ylim2, w, dtStart)

Might be easier to see correlation in a dot plot format.

## Warning: Removed 32 rows containing non-finite values (stat_smooth).

This is a legacy series from FRED. It has not been updated in a couple of years so I am assuming it will go away.

Federal Reserve Reverse Repo Agreements

Compare liabilities to reverse repo trends

datay <- "WLRRAL"
ylim <- c(0, 700)
dtStart = as.Date('2003-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dtStart)

Spiky, might be easier to look at year-over-year

datay <- "WLRRAL_YoY"
ylim <- c(-100, 100)
dtStart = as.Date('2003-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dtStart)

Normalized by GDP

datay <- "WLRRAL.by.GDP"
ylim <- c(0, 4)
dtStart = as.Date('2003-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dtStart)

Overnight Bank Funding Rate

“The overnight bank funding rate is calculated using federal funds transactions and certain Eurodollar transactions. The federal funds market consists of domestic unsecured borrowings in U.S. dollars by depository institutions from other depository institutions and certain other entities, primarily government-sponsored enterprises, while the Eurodollar market consists of unsecured U.S. dollar deposits held at banks or bank branches outside of the United States. U.S.-based banks can also take Eurodollar deposits domestically through international banking facilities (IBFs). The overnight bank funding rate (OBFR) is calculated as a volume-weighted median of overnight federal funds transactions and Eurodollar transactions reported in the FR 2420 Report of Selected Money Market Rates. Volume-weighted median is the rate associated with transactions at the 50th percentile of transaction volume. Specifically, the volume-weighted median rate is calculated by ordering the transactions from lowest to highest rate, taking the cumulative sum of volumes of these transactions, and identifying the rate associated with the trades at the 50th percentile of dollar volume. The published rates are the volume-weighted median transacted rate, rounded to the nearest basis point.” https://www.newyorkfed.org/markets/obfrinfo.

Secured Overnight Financing Rate

“The Secured Overnight Financing Rate (SOFR) is a broad measure of the cost of borrowing cash overnight collateralized by Treasury securities. The SOFR includes all trades in the Broad General Collateral Rate plus bilateral Treasury repurchase agreement (repo) transactions cleared through the Delivery-versus-Payment (DVP) service offered by the Fixed Income Clearing Corporation (FICC), which is filtered to remove a portion of transactions considered “specials” " https://apps.newyorkfed.org/markets/autorates/sofr

Take a look at the variation (99th - 1st percentile)

Reserve Balances with Federal Reserve Banks

Hard to get a sense of these series in the absolute. Take a look relative to GDP.

Correlation Between Reserves and Total Loans

As reserves increase there should be less lending. That correlation generally holds.

## Scale for 'y' is already present. Adding another scale for 'y', which
## will replace the existing scale.

Did the reserve balances increase after the 2016 and 2018 drops? Not in the same way. There are some relationships between the equities market and the reserves though.

datay <- "WRESBAL_YoY"
ylim <- c(-40, 20)
dtStart = as.Date('2015-01-01')
p1 <- plotSingleQuick(dfRecession, df.data, datay, ylim, dtStart)
p1 <-
  p1 + geom_vline(
    xintercept = as.Date("2015-08-24"),
    linetype = "dashed",
    color = "grey",
    size = 1.0
  )
p1 <-
  p1 + geom_vline(
    xintercept = as.Date("2016-01-08"),
    linetype = "dashed",
    color = "grey",
    size = 1.0
  )
p1 <-
  p1 + geom_vline(
    xintercept = as.Date("2018-02-05"),
    linetype = "dashed",
    color = "grey",
    size = 1.0
  )
p1 <-
  p1 + geom_vline(
    xintercept = as.Date("2018-10-11"),
    linetype = "dashed",
    color = "grey",
    size = 1.0
  )

datay <- "GSPC.Open"
datay_aux <- "GSPC.Close"
ylim <- c(1500, 3200)
p2 <-
  plotSingle(
    dfRecession,
    df.data,
    "date",
    datay,
    getPlotTitle(df.symbols, datay),
    "Date",
    getPlotYLabel(df.symbols, datay),
    c(dtStart, Sys.Date()),
    ylim,
    b.legend = FALSE,
    b.percentile = FALSE,
    b.long.legend = FALSE
  )

p2 <-
  p2 + geom_vline(
    xintercept = as.Date("2015-08-24"),
    linetype = "dashed",
    color = "grey",
    size = 1.0
  )
p2 <-
  p2 + geom_vline(
    xintercept = as.Date("2016-01-08"),
    linetype = "dashed",
    color = "grey",
    size = 1.0
  )
p2 <-
  p2 + geom_vline(
    xintercept = as.Date("2018-02-05"),
    linetype = "dashed",
    color = "grey",
    size = 1.0
  )
p2 <-
  p2 + geom_vline(
    xintercept = as.Date("2018-10-11"),
    linetype = "dashed",
    color = "grey",
    size = 1.0
  )


grid.arrange(p1,
             p2,
             ncol = 1,
             top = "Reserve balance YoY change and S&P Price")

Explicitly correlate reserve balances and total loans. It is a weak and noisy correlation.

## Warning: Removed 428 rows containing non-finite values (stat_smooth).

Interest on excess reserves

Money supplies

Basic currency trend (currency component of M1)

datay <- "WCURRNS"
dtStart = as.Date('1980-01-01')
ylim <- c(0, 1800)
myplot <- plotSingleQuick(dfRecession, df.data, datay, ylim, dtStart)
myplot

datay <- "WCURRNS_YoY"
dtStart = as.Date('1980-01-01')
ylim <- c(0, 15)
myplot <- plotSingleQuick(dfRecession, df.data, datay, ylim, dtStart)
myplot

datay <- "WCURRNS_YoY"
dtStart = as.Date('2000-01-01')
ylim <- c(0, 15)
myplot <- plotSingleQuick(dfRecession, df.data, datay, ylim, dtStart)
myplot

The rate of change of money supply could be an indicator of a recession. Let’s see how that compares.

datay <- "M1_YoY"
ylim <- c(-15, 25)
datay_aux <- "M2_YoY"
dtStart = as.Date('1980-01-01')
myplot <- plotSingleQuick(dfRecession, df.data, datay, ylim, dtStart)

myPlot <- plotSingle(dfRecession, df.data, "date", datay, getPlotTitle(df.symbols, datay), "Date", 
            getPlotYLabel(df.symbols, datay), c(dtStart, Sys.Date()), ylim, TRUE)
myPlot + geom_line(data=df.data, aes_string(x="date", y=datay_aux, colour=shQuote(datay_aux)), na.rm = TRUE)

Intervention in the repo market

The federal reserve provides liquidity to the repo market, summary of that action

Federal debt as percent GDP

datay <- "GFDEGDQ188S"
ylim <- c(30, 110)
plotSingleQuick(dfRecession, df.data, datay, ylim)

Federal deficit as percent GDP

datay <- "FYFSGDA188S"
ylim <- c(-30, 5)
plotSingleQuick(dfRecession, df.data, datay, ylim)

Charlie Hatch has a nice format of deficit versus debt:

## Scale for 'y' is already present. Adding another scale for 'y', which
## will replace the existing scale.

Total loans

One business cycle theory describes recessions as a market adjustment to mis-allocated assets, often fueled by an credit expansion. That makes the volume of loans an interesting feature to look at. In the presentation of data it looks like the great recession had the largest impact.

Plotting the year over year growth rate helps pull out those small changes in the early years in the data. Peaks can be seen prior to most recessions.

Zoom in to the last couple of decades

As long term interest rates rise, loans should start to tick down. To check this, the total loans and 10 to 1 year spreads are plotted. This is generally the trend observed.

There is a good correlation between these two variables. This next section plots that correction explicitly.

Total loans as percent of GDP

This is the total loans. I think the picture is too broad to point to a specific sector of the economy.

Commercial and industral loans

Business loans should slow before the recession (a contraction in credit as rates rise).

Commercial and industrial loans as percent of GDP and and income

Look at business debt normalized by GDP over the entire time series

Farm loans

See how the farming sector is fairing.

Real estate loans

Data taken from H.8 Assets and Liabilities of Commercial Banks in the United States. Take a look at SA and NSA data series as weekly and month updates. It should all be similar at this scale.

This gives a big picture, but makes it hard to connect the loans with the income needed to cover those loans. In the next section, loans will be broken up by commercial and residential.

Real Estate (Residential)

In absolute terms the mortgages have increased, but it does not appear to be out of line with the overall economy.

Normalized by GDP it is easier to see the peak in 2008 and that loan levels appear reasonable at the commercial banks.

Maybe the GSE’s are making loans. Take a look at the total mortgages from Z.1 as a percentage of GDP. That does not look too far off trend (ignoring that peak in 2008).

I am assuming that personal income is paying for the mortgages.

Real estate (residential) as percent of GDP and and income

## Warning: Removed 1 rows containing missing values (geom_text).

## Warning: Removed 1 rows containing missing values (geom_text).

Consumer loans

Focusing on the consumer sector the growth in debt and incomes can be directly compared. Personal income, as a percent of GDP, remains nearly constant. It is not uncommon for the personal income to rise prior to a recession. Likely this reflect increasing asset prices and market returns. Also interesting to see the loans pick up after interest rates dropped in 1982.

Consumer loans as percent of GDP and and income

## Warning: Removed 1 rows containing missing values (geom_text).

## Warning: Removed 1 rows containing missing values (geom_text).

Take a closer look since the 2008 recession. Looks like loans are starting to slow as the interest burden rises and incomes remain stable. There are some anomolies in the A065RC1A027NBEA data series because it only updates onces a year. the PI series updates once a month but is noisier and seasonally adjusted. It also shows incomes rising in the middle of the 2008 recession, which doesn’t seem to be accurate.

## Warning: Removed 1 rows containing missing values (geom_text).

## Warning: Removed 1 rows containing missing values (geom_text).

Nonfincial corporate business security repo asset level

T-Bills and Yield Curve

Speaking of loans, interest rates also play into this. This analysis will focus on treasure bills. The 3-month is plotted below. The yield flattens before a recession as investors go long on bonds and short on equities.

datay <- "TB3MS"
datay.aux <- "DTB3"
ylim <- c(0, 20)
p1 <- plotSingleQuickModern(datay, ylim)
p1 + geom_line(data=df.data, aes_string(x="date", y=datay.aux, colour=shQuote(datay.aux)), na.rm = TRUE)

datay <- "TB3MS"
datay.aux <- "DTB3"
ylim <- c(0, 2.5)
dtStart = as.Date('2017-01-01')
p1 <- plotSingle(dfRecession, df.data, "date", datay, getPlotTitle(df.symbols, datay), "Date", 
            getPlotYLabel(df.symbols, datay), c(dtStart, Sys.Date()), ylim, TRUE)
p1 + geom_line(data=df.data, aes_string(x="date", y=datay.aux, colour=shQuote(datay.aux)), na.rm = TRUE)

Check out LIBOR and fed funds rate

datay <- "TB3MS"
datay_aux <- "USD1MTD156N"
ylim <- c(0, 12)
dtStart = as.Date('1985-01-01')
myPlot <- plotSingle(dfRecession, df.data, "date", datay, getPlotTitle(df.symbols, datay), "Date", 
            getPlotYLabel(df.symbols, datay), c(dtStart, Sys.Date()), ylim, TRUE)
myPlot <- myPlot + geom_line(data=df.data, aes_string(x="date", y=datay_aux, colour=shQuote(datay_aux)), na.rm = TRUE)

myPlot

The 1-year is plotted below. The yield flattens before a recession as investors go long on bonds and short on equities.

datay <- "DGS1"
ylim <- c(0, 20)
p1 <- plotSingleQuickModern(datay, ylim)
p1

datay <- "DGS10"
datay.aux <- "TNX.Close"
ylim <- c(0, 20)
p1 <- plotSingleQuickModern(datay, ylim)
p1 + geom_line(data=df.data, aes_string(x="date", y=datay.aux, colour=shQuote(datay.aux)), na.rm = TRUE)

Close in, the trend towards inversion be more easily seen. I am also comparing data from the CBOE as well as FRED.

Bond yields are a good proxy for interest rates. As rates rise the theory goes that loans should decrease (inverse correlation).

And a longer window

The yield curve (30 year bond rate minus the 10 year bond rate) may not be a good recession indicator, but a collapse is not good (https://blogs.wsj.com/moneybeat/2018/04/30/theres-more-than-one-part-of-the-yield-curve-getting-flatter/).

The yield curve (10 year bond rate minus the 1 year bond rate) seems to a good indicator of an oncoming recession. It could be a buy indicator by itself.

More recent data

Just the last 24 months or so.

Plot the 10 Year to 3 month over a few decades to see what the outling cases look like

The last two year compare favorably with the period around the 2015-2016 turndown, driven primarily by slowing of the Chinese GDP. Not a debt-driven cycle.

This plot format was suggested by a mises.org article (https://mises.org/wire/yield-curve-accordion-theory), but they only went back to 1988. The date seemed arbitrary so I went back further in time.

Take a look at more recent data

Try looking at a 1-year average of the above time series

High quality bonds

datay <- "AAA"
ylim <- c(2.5, 10)
dtStart = as.Date('1997-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dtStart)

High quality bonds to 10-year treasury

datay <- "DGS10ByAAA"
ylim <- c(1, 2.5)
dtStart = as.Date('1967-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dtStart)

High yield spread

“This data represents the Option-Adjusted Spread (OAS) of the ICE BofAML US Corporate A Index, a subset of the ICE BofAML US Corporate Master Index tracking the performance of US dollar denominated investment grade rated corporate debt publically issued in the US domestic market. This subset includes all securities with a given investment grade rating A. The ICE BofAML OASs are the calculated spreads between a computed OAS index of all bonds in a given rating category and a spot Treasury curve. An OAS index is constructed using each constituent bond‚Äôs OAS, weighted by market capitalization. When the last calendar day of the month takes place on the weekend, weekend observations will occur as a result of month ending accrued interest adjustments.”

• ICE Benchmark Administration Limited (IBA), ICE BofAML US Corporate A Option-Adjusted Spread [BAMLC0A3CA], retrieved from FRED, Federal Reserve Bank of St. Louis; https://fred.stlouisfed.org/series/BAMLC0A3CA, July 4, 2019.

datay <- "BAMLC0A3CA"
ylim <- c(0, 7)
dtStart = as.Date('1997-01-01')
plotSingleQuick(dfRecession, df.data, datay, ylim, dtStart)

Municipal bond market

Suggest by a WSJ article, change in volume for high-risk muni’s. Doesn’t look like there is much too it yet.

https://www.wsj.com/articles/risky-municipal-bonds-are-on-a-hot-streak-11558949401?mod=hp_lead_pos3

datay <- "HYMB.Close"
ylim <- c(55, 60)
dtStart = as.Date('2015-01-01')
p1 <- plotSingleQuick(dfRecession, df.data, datay, ylim, dtStart)

p1 <-
  p1 + geom_vline(
    xintercept = as.Date("2015-08-24"),
    linetype = "dashed",
    color = "grey",
    size = 1.0
  )

p1 <-
  p1 + geom_vline(
    xintercept = as.Date("2016-01-08"),
    linetype = "dashed",
    color = "grey",
    size = 1.0
  )
p1 <-
  p1 + geom_vline(
    xintercept = as.Date("2018-02-05"),
    linetype = "dashed",
    color = "grey",
    size = 1.0
  )
p1 <-
  p1 + geom_vline(
    xintercept = as.Date("2018-10-11"),
    linetype = "dashed",
    color = "grey",
    size = 1.0
  )

datay <- "HYMB.Volume"
ylim <- c(0, 1200000)
dtStart = as.Date('2015-01-01')
p1.vol <- plotSingleQuick(dfRecession, df.data, datay, ylim, dtStart)

p1.vol <-
  p1.vol + geom_vline(
    xintercept = as.Date("2015-08-24"),
    linetype = "dashed",
    color = "grey",
    size = 1.0
  )

p1.vol <-
  p1.vol + geom_vline(
    xintercept = as.Date("2016-01-08"),
    linetype = "dashed",
    color = "grey",
    size = 1.0
  )
p1.vol <-
  p1.vol + geom_vline(
    xintercept = as.Date("2018-02-05"),
    linetype = "dashed",
    color = "grey",
    size = 1.0
  )
p1.vol <-
  p1.vol + geom_vline(
    xintercept = as.Date("2018-10-11"),
    linetype = "dashed",
    color = "grey",
    size = 1.0
  )


datay <- "GSPC.Open"
datay_aux <- "GSPC.Close"
ylim <- c(1500, 3000)
p2 <-
  plotSingle(
    dfRecession,
    df.data,
    "date",
    datay,
    getPlotTitle(df.symbols, datay),
    "Date",
    getPlotYLabel(df.symbols, datay),
    c(dtStart, Sys.Date()),
    ylim,
    TRUE
  )

p2 <-
  p2 + geom_vline(
    xintercept = as.Date("2015-08-24"),
    linetype = "dashed",
    color = "grey",
    size = 1.0
  )
p2 <-
  p2 + geom_vline(
    xintercept = as.Date("2016-01-08"),
    linetype = "dashed",
    color = "grey",
    size = 1.0
  )
p2 <-
  p2 + geom_vline(
    xintercept = as.Date("2018-02-05"),
    linetype = "dashed",
    color = "grey",
    size = 1.0
  )
p2 <-
  p2 + geom_vline(
    xintercept = as.Date("2018-10-11"),
    linetype = "dashed",
    color = "grey",
    size = 1.0
  )


grid.arrange(p1,
             p1.vol,
             p2,
             ncol = 1,
             top = "High Yield Muni's and S&P Price")

Total Loans and yield curve correlation

This relationship was suggest by Charlie and it is an interesting one. As the yield curve flattens (10-year and 1-year rates converge), total loans grow. The generalization is not always accurate, but it does fit.

I wanted to see how this looked compared to the 3 month

Consumer loans and yield curve correlation

Compared to business loans, consumer loans seem to have to response to the 10Y to 3M yield curve.

Business loans and yield curve correlation

That’s pretty good correlation. Let’s see what the rolling correlation looks like.

datay1 <- "TOTLNNSA_YoY"
ylim1 <- c(-10, 20)

datay2 <- "DGS10TO1"
ylim2 <- c(-5, 10)

dtStart <- as.Date("1jan1960","%d%b%Y")

w <- 360
corrName <- calcRollingCorr(dfRecession, df.data, df.symbols, datay1, ylim1, datay2, ylim2, w, dtStart)

datay1 <- "TOTLNNSA_YoY"
ylim1 <- c(-10, 20)

datay2 <- "DGS10TO1"
ylim2 <- c(-5, 10)

dtStart <- as.Date("1jan1960","%d%b%Y")

w <- 720
corrName <- calcRollingCorr(dfRecession, df.data, df.symbols, datay1, ylim1, datay2, ylim2, w, dtStart)

One other items, let’s see how loans do versus the federal funds rate

TAble 6 - Light Vehicle and Total Vehicle Sales

Table 1.1.5. Gross Domestic Product

[Billions of dollars] Seasonally adjusted at annual rates

Table 1.1.9. Implicit Price Deflators for Gross Domestic Product

[Index numbers, 2012=100] Seasonally adjusted

GDP normalized by CPI

Normalize GDP by CPI

Economic yield curve (GDP to 1-year treasury)

GDP versus the yield on the 1-year. This series was prompted by an article suggesting that the “economic yield curve” should be used to indicate a recession rather than an inverted yield curve. Less of indicator and more of concurrent confirmation of recession. Not sure why they would be related either.

Economic yield curve (GDP to 3-month treasury)

Same idea as above, but applied the 3-month treasury.This one has fewer false triggers, but is not as helpful as 10Y to 3M spread in predicting a recession.

Table 6.16D. Corporate Profits by Industry

Select series from Table 6.16D

Table 2.6. Personal Income and Its Disposition, Monthly

Billions of dollars; months are seasonally adjusted at annual rates.

datay1 <- "UNRATE"
ylim1 <- c(2, 12)

datay2 <- "PSAVERT"
ylim2 <- c(0, 15)

dtStart <- as.Date("1jan1985","%d%b%Y")

w <- 360
corrName <- calcRollingCorr(dfRecession, df.data, df.symbols, datay1, ylim1, datay2, ylim2, w, dtStart)

Personal savings to household net worth

A relationship between personal savings and household networth can be seen in a scatter plot. This was suggested by a WSJ article (https://blogs.wsj.com/dailyshot/2018/02/23/the-daily-shot-reasons-for-declining-u-s-household-savings-rate/).

U.S. International Trade in Goods and Services (FT900)

U.S. Bureau of Economic Analysis and U.S. Census Bureau, U.S. Imports of Goods by Customs Basis from China [IMPCH], retrieved from FRED, Federal Reserve Bank of St. Louis; https://fred.stlouisfed.org/series/IMPCH, October 5, 2019.

New Houses Sold and For Sale by Stage of Construction and Median Number of Months on Sales Market

Read an article suggesting that housing sales and sales growth could be useful. FRED only has new home data so start there.

datay <- "HSN1FNSA"
ylim <- c(0, 200)
dtStart = as.Date('1964-01-01')
p1 <- plotSingleQuick(dfRecession, df.data, datay, ylim, dtStart)

datay <- "HNFSUSNSA"
ylim <- c(0, 600)
p2 <- plotSingleQuick(dfRecession, df.data, datay, ylim, dtStart)

datay <- "HNFSUSNSA.minus.HSN1FNSA"
ylim <- c(0, 600)
p3 <-
  plotSingle(
    dfRecession,
    df.data,
    "date",
    datay,
    getPlotTitle(df.symbols, datay),
    "Date",
    getPlotYLabel(df.symbols, datay),
    c(dtStart, Sys.Date()),
    ylim,
    TRUE
  )

grid.arrange(p1,
             p2,
             p3,
             ncol = 1,
             top = "New Housing Sales")

New housing yoy

datay <- "HSN1FNSA_YoY"
ylim <- c(-100, 50)
dtStart = as.Date('1964-01-01')
p1 <- plotSingleQuick(dfRecession, df.data, datay, ylim, dtStart)

datay <- "HNFSUSNSA_YoY"
p2 <- plotSingleQuick(dfRecession, df.data, datay, ylim, dtStart)

datay <- "HNFSUSNSA.minus.HSN1FNSA_YoY"
p3 <-
  plotSingle(
    dfRecession,
    df.data,
    "date",
    datay,
    getPlotTitle(df.symbols, datay),
    "Date",
    getPlotYLabel(df.symbols, datay),
    c(dtStart, Sys.Date()),
    ylim,
    TRUE
  )

grid.arrange(p1,
             p2,
             p3,
             ncol = 1,
             top = "New Housing Sales Year-Over-Year")

Page 4: Not Seasonally adjusted, billions of dollars

L.102 Nonfinancial Business

L.214 Loans

B.101 Balance Sheet of Households and Nonprofit Organizations

Population to GDP

Look at GDP divided by CPI per person. It flattens and even dips a little prior to a recession. Might be worth looking at the derivative of this series.

That is worth a closer look

datay1 <- "GDPBYCPIAUCSLBYPOPTHM_SmoothDer"
ylim1 <- c(-5, 5)

datay2 <- "RecInit_Smooth"
ylim2 <- c(0, 1)

dtStart <- as.Date("1jan1960","%d%b%Y")

w <- 30
corrName <- calcRollingCorr(dfRecession, df.data, df.symbols, datay1, ylim1, datay2, ylim2, w, dtStart)